Microsoft AI: MAI-Image-2.6 Flash

Microsoft AI: MAI-Image-2.6 Flash is an image generation model from Microsoft, part of the MAI-Image-2.6 family. Within that family it is positioned as the lower-latency, lower-cost member, built for latency-sensitive, high-throughput production work. It is aimed at teams that need image generation and editing at scale rather than one-off, carefully tuned outputs.

The model accepts text and image inputs and produces images as its only output. Its context window is 4096 tokens, and the maximum output is 1024. Generation is controlled through a small set of image parameters: aspect_ratio, which accepts 1:1, 4:3, 3:4, 16:9, 9:16, 3:2, 2:3 or auto; n, which is fixed at 1 image per request; and input_references, which accepts between 0 and 5 reference images. Alongside these, the model supports max_completion_tokens, max_tokens and temperature.

Those capabilities point to a fairly specific set of uses. Because it takes image inputs as well as text, it fits editing and reference-driven generation, where you supply up to five images and describe the change or the result you want. The aspect ratio options cover square, landscape and portrait formats, so the same model can serve social, web and print-oriented layouts without a separate pipeline. The combination of image input and image output also suits batch workflows such as generating variants from a product photo, restyling assets, or producing placeholder and draft imagery inside an automated build. The Flash positioning suggests it is the right pick when throughput and response time matter more than the last increment of precision.

You call it through api14's OpenAI-compatible API using the model id "microsoft/mai-image-2.6-flash". Requests follow the same shape as other OpenAI-compatible chat completions calls, with your text prompt and any reference images in the input and the generated image returned in the response. The supported request parameters are max_completion_tokens, max_tokens and temperature, plus the image parameters aspect_ratio, n and input_references. Set n to 1, choose an aspect ratio or leave it as auto, and pass up to five reference images when you want the output grounded in existing visuals.

A few limits are worth keeping in mind. The context window is 4096 tokens, which is modest, so very long prompts or many large reference images may need trimming. Only one image is returned per request, so producing a set of variants means multiple calls. If your work depends on maximum precision rather than latency and throughput, the precision tier of the same family is the more natural fit.